Spiking Pseudo-Ensembles Boost OOD Detection in Remote Sensing.

Spiking Neural Networks (SNNs) are highly appealing for remote sensing applications due to their energy efficiency, especially in resource-constrained environments. However, ensuring reliable detection of out-of-distribution (OOD) data remains a significant challenge for SNNs. While deep ensembles offer strong predictive uncertainty, they demand substantial computational resources by requiring multiple full models. This research proposes an efficient alternative: a spiking pseudo-ensemble. This pseudo-ensemble attaches multiple lightweight classification heads to a single, frozen SNN backbone. A key innovation is the introduction of an "agree-disagree" objective during training. This objective prevents "diversity collapse," a common issue where independently parameterized heads produce correlated predictions. Instead, it encourages the heads to agree on clean in-distribution samples while promoting diverse predictions on structured, uncertainty-inducing transformations of those same inputs. This method provides a diversity-promoting signal without needing external OOD data. Experiments show that this approach matches or surpasses the performance of traditional deep ensembles with significantly fewer parameters and backbone evaluations, making OOD detection more practical for edge devices.